Spaces:
Running
on
Zero
Running
on
Zero
Update app.py
Browse files
app.py
CHANGED
@@ -18,6 +18,9 @@ latent_scale_factor = 0.18215 # Same as in DiTTrainer
|
|
18 |
# For tracking progress in UI
|
19 |
global_progress = 0
|
20 |
|
|
|
|
|
|
|
21 |
def load_dit_model(dit_size):
|
22 |
"""Load DiT model of specified size"""
|
23 |
#ckpt_path = f"./ckpts/DiT_{dit_size}_final.pth"
|
@@ -45,8 +48,9 @@ def load_dit_model(dit_size):
|
|
45 |
return model
|
46 |
|
47 |
class DiffusionSampler:
|
48 |
-
def __init__(self, device="cuda" if torch.cuda.is_available() else "cpu"):
|
49 |
self.device = device
|
|
|
50 |
self.vae = None
|
51 |
|
52 |
# Pre-compute diffusion parameters
|
@@ -65,6 +69,14 @@ class DiffusionSampler:
|
|
65 |
self.sqrt_recip_alphas = self.sqrt_recip_alphas.to(self.device)
|
66 |
self.betas = self.betas.to(self.device)
|
67 |
self.posterior_variance = self.posterior_variance.to(self.device)
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
68 |
|
69 |
def load_vae(self):
|
70 |
"""Load VAE model (done lazily to save memory until needed)"""
|
@@ -99,6 +111,8 @@ class DiffusionSampler:
|
|
99 |
|
100 |
# Start with random latents
|
101 |
latents = torch.randn((num_samples, 4, 32, 32), device=self.device)
|
|
|
|
|
102 |
|
103 |
# Use classifier-free guidance for better quality
|
104 |
cfg_scale = 2.5
|
@@ -135,6 +149,10 @@ class DiffusionSampler:
|
|
135 |
|
136 |
# Decode latents to images
|
137 |
self.load_vae()
|
|
|
|
|
|
|
|
|
138 |
latents = latents / self.vae.config.scaling_factor
|
139 |
latents = latents.to(self.device)
|
140 |
|
|
|
18 |
# For tracking progress in UI
|
19 |
global_progress = 0
|
20 |
|
21 |
+
# Enable half precision inference
|
22 |
+
USE_HALF_PRECISION = True
|
23 |
+
|
24 |
def load_dit_model(dit_size):
|
25 |
"""Load DiT model of specified size"""
|
26 |
#ckpt_path = f"./ckpts/DiT_{dit_size}_final.pth"
|
|
|
48 |
return model
|
49 |
|
50 |
class DiffusionSampler:
|
51 |
+
def __init__(self, device="cuda" if torch.cuda.is_available() else "cpu", use_half = USE_HALF_PRECISION):
|
52 |
self.device = device
|
53 |
+
self.use_half = use_half
|
54 |
self.vae = None
|
55 |
|
56 |
# Pre-compute diffusion parameters
|
|
|
69 |
self.sqrt_recip_alphas = self.sqrt_recip_alphas.to(self.device)
|
70 |
self.betas = self.betas.to(self.device)
|
71 |
self.posterior_variance = self.posterior_variance.to(self.device)
|
72 |
+
|
73 |
+
# Convert to half precision if needed
|
74 |
+
if self.use_half:
|
75 |
+
self.sqrt_alphas_cumprod = self.sqrt_alphas_cumprod.half()
|
76 |
+
self.sqrt_one_minus_alpha_cumprod = self.sqrt_one_minus_alpha_cumprod.half()
|
77 |
+
self.sqrt_recip_alphas = self.sqrt_recip_alphas.half()
|
78 |
+
self.betas = self.betas.half()
|
79 |
+
self.posterior_variance = self.posterior_variance.half()
|
80 |
|
81 |
def load_vae(self):
|
82 |
"""Load VAE model (done lazily to save memory until needed)"""
|
|
|
111 |
|
112 |
# Start with random latents
|
113 |
latents = torch.randn((num_samples, 4, 32, 32), device=self.device)
|
114 |
+
if self.use_half:
|
115 |
+
latents = latents.half()
|
116 |
|
117 |
# Use classifier-free guidance for better quality
|
118 |
cfg_scale = 2.5
|
|
|
149 |
|
150 |
# Decode latents to images
|
151 |
self.load_vae()
|
152 |
+
|
153 |
+
# Convert back to float
|
154 |
+
if self.use_half:
|
155 |
+
latents = latents.float()
|
156 |
latents = latents / self.vae.config.scaling_factor
|
157 |
latents = latents.to(self.device)
|
158 |
|